Efficient Privacy-Preserving Prediction for Three-Layer Feedforward Neural Networks Using Ring-LWE-based Homomorphic Encryption

Abstract

Concerns about privacy of data prevent from making good use of a huge amount of data. Data analysis while preserving privacy is a very important task. In this research, we propose a Privacy-Preserving Machine Learning that can efficiently compute inner product in a three-layered neural network using Ring-LWE-based Homomorphic Encryption. We propose a two-party model consisting of client and server: the former encrypts input data and receives a classification result from a server and the latter performs predicting process over the encrypted data using a trained classification model. This enables that the client acquires the inference result without revealing the privacy of their data and the server protects their model from exposing it. The proposed method costs 10.549 [ms] per one class for prediction process and performed keeping its accuracy close to the case of sigmoid and ReLU.